[SKX] avx512_icmp_packed multiclass extension
authorRobert Khasanov <rob.khasanov@gmail.com>
Mon, 25 Aug 2014 14:49:34 +0000 (14:49 +0000)
committerRobert Khasanov <rob.khasanov@gmail.com>
Mon, 25 Aug 2014 14:49:34 +0000 (14:49 +0000)
Extended avx512_icmp_packed multiclass by masking versions.
Added avx512_icmp_packed_rmb multiclass for embedded broadcast versions.
Added corresponding _vl multiclasses.
Added encoding tests for CPCMP{EQ|GT}* instructions.
Add more fields for X86VectorVTInfo.
Added AVX512VLVectorVTInfo that include X86VectorVTInfo for 512/256/128-bit versions

Differential Revision: http://reviews.llvm.org/D5024

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@216383 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/X86/X86InstrAVX512.td
test/MC/X86/avx512-encodings.s
test/MC/X86/x86-64-avx512bw.s
test/MC/X86/x86-64-avx512bw_vl.s
test/MC/X86/x86-64-avx512f_vl.s
utils/TableGen/X86DisassemblerTables.cpp
utils/TableGen/X86RecognizableInstr.cpp

index 2b84b6c..a372a60 100644 (file)
@@ -22,34 +22,83 @@ class X86VectorVTInfo<int NumElts, ValueType EltVT, RegisterClass rc,
   // Suffix used in the instruction mnemonic.
   string Suffix = suffix;
 
+  string VTName = "v" # NumElts # EltVT;
+
   // The vector VT.
-  ValueType VT = !cast<ValueType>("v" # NumElts # EltVT);
+  ValueType VT = !cast<ValueType>(VTName);
 
   string EltTypeName = !cast<string>(EltVT);
   // Size of the element type in bits, e.g. 32 for v16i32.
-  string EltSize = !subst("i", "", !subst("f", "", EltTypeName));
+  string EltSizeName = !subst("i", "", !subst("f", "", EltTypeName));
+  int EltSize = EltVT.Size;
 
   // "i" for integer types and "f" for floating-point types
-  string TypeVariantName = !subst(EltSize, "", EltTypeName);
+  string TypeVariantName = !subst(EltSizeName, "", EltTypeName);
 
   // Size of RC in bits, e.g. 512 for VR512.
   int Size = VT.Size;
 
   // The corresponding memory operand, e.g. i512mem for VR512.
   X86MemOperand MemOp = !cast<X86MemOperand>(TypeVariantName # Size # "mem");
+  X86MemOperand ScalarMemOp = !cast<X86MemOperand>(EltVT # "mem");
+
+  // Load patterns
+  // Note: For 128/256-bit integer VT we choose loadv2i64/loadv4i64
+  //       due to load promotion during legalization
+  PatFrag LdFrag = !cast<PatFrag>("load" #
+                                  !if (!eq (TypeVariantName, "i"),
+                                       !if (!eq (Size, 128), "v2i64",
+                                       !if (!eq (Size, 256), "v4i64",
+                                            VTName)), VTName));
+  PatFrag ScalarLdFrag = !cast<PatFrag>("load" # EltVT);
 
   // The corresponding float type, e.g. v16f32 for v16i32
-  ValueType FloatVT = !if (!eq(TypeVariantName, "i"),
-                           !cast<ValueType>("v" # NumElts # "f" # EltSize),
-                           VT);
+  // Note: For EltSize < 32, FloatVT is illegal and TableGen
+  //       fails to compile, so we choose FloatVT = VT
+  ValueType FloatVT = !cast<ValueType>(
+                        !if (!eq (!srl(EltSize,5),0),
+                             VTName,
+                             !if (!eq(TypeVariantName, "i"),
+                                  "v" # NumElts # "f" # EltSize,
+                                  VTName)));
 
   // The string to specify embedded broadcast in assembly.
   string BroadcastStr = "{1to" # NumElts # "}";
 }
 
+def v64i8_info  : X86VectorVTInfo<64,  i8, VR512, "b">;
+def v32i16_info : X86VectorVTInfo<32, i16, VR512, "w">;
 def v16i32_info : X86VectorVTInfo<16, i32, VR512, "d">;
 def v8i64_info  : X86VectorVTInfo<8,  i64, VR512, "q">;
 
+// "x" in v32i8x_info means RC = VR256X
+def v32i8x_info  : X86VectorVTInfo<32,  i8, VR256X, "b">;
+def v16i16x_info : X86VectorVTInfo<16, i16, VR256X, "w">;
+def v8i32x_info  : X86VectorVTInfo<8,  i32, VR256X, "d">;
+def v4i64x_info  : X86VectorVTInfo<4,  i64, VR256X, "q">;
+
+def v16i8x_info  : X86VectorVTInfo<16,  i8, VR128X, "b">;
+def v8i16x_info  : X86VectorVTInfo<8,  i16, VR128X, "w">;
+def v4i32x_info  : X86VectorVTInfo<4,  i32, VR128X, "d">;
+def v2i64x_info  : X86VectorVTInfo<2,  i64, VR128X, "q">;
+
+class AVX512VLVectorVTInfo<X86VectorVTInfo i512, X86VectorVTInfo i256,
+                           X86VectorVTInfo i128> {
+  X86VectorVTInfo info512 = i512;
+  X86VectorVTInfo info256 = i256;
+  X86VectorVTInfo info128 = i128;
+}
+
+def avx512vl_i8_info  : AVX512VLVectorVTInfo<v64i8_info, v32i8x_info,
+                                             v16i8x_info>;
+def avx512vl_i16_info : AVX512VLVectorVTInfo<v32i16_info, v16i16x_info,
+                                             v8i16x_info>;
+def avx512vl_i32_info : AVX512VLVectorVTInfo<v16i32_info, v8i32x_info,
+                                             v4i32x_info>;
+def avx512vl_i64_info : AVX512VLVectorVTInfo<v8i64_info, v4i64x_info,
+                                             v2i64x_info>;
+
+
 // Common base class of AVX512_masking and AVX512_masking_3src.
 multiclass AVX512_masking_common<bits<8> O, Format F, dag Outs, dag Ins,
                                  dag MaskingIns, dag ZeroMaskingIns,
@@ -971,34 +1020,131 @@ defm VCMPSDZ : avx512_cmp_scalar<FR64X, f64mem, AVXCC, X86cmpms, f64, loadf64,
                  XD, VEX_W;
 }
 
-multiclass avx512_icmp_packed<bits<8> opc, string OpcodeStr, RegisterClass KRC, 
-              RegisterClass RC, X86MemOperand x86memop, PatFrag memop_frag, 
-              SDNode OpNode, ValueType vt> {
+multiclass avx512_icmp_packed<bits<8> opc, string OpcodeStr, SDNode OpNode,
+              X86VectorVTInfo _> {
   def rr : AVX512BI<opc, MRMSrcReg,
-             (outs KRC:$dst), (ins RC:$src1, RC:$src2), 
-             !strconcat(OpcodeStr, " \t{$src2, $src1, $dst|$dst, $src1, $src2}"),
-             [(set KRC:$dst, (OpNode (vt RC:$src1), (vt RC:$src2)))], 
+             (outs _.KRC:$dst), (ins _.RC:$src1, _.RC:$src2),
+             !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
+             [(set _.KRC:$dst, (OpNode (_.VT _.RC:$src1), (_.VT _.RC:$src2)))],
              IIC_SSE_ALU_F32P_RR>, EVEX_4V;
+  let mayLoad = 1 in
   def rm : AVX512BI<opc, MRMSrcMem,
-             (outs KRC:$dst), (ins RC:$src1, x86memop:$src2), 
-             !strconcat(OpcodeStr, " \t{$src2, $src1, $dst|$dst, $src1, $src2}"),
-             [(set KRC:$dst, (OpNode (vt RC:$src1), (memop_frag addr:$src2)))],
+             (outs _.KRC:$dst), (ins _.RC:$src1, _.MemOp:$src2),
+             !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
+             [(set _.KRC:$dst, (OpNode (_.VT _.RC:$src1),
+                                     (_.VT (bitconvert (_.LdFrag addr:$src2)))))],
              IIC_SSE_ALU_F32P_RM>, EVEX_4V;
+  def rrk : AVX512BI<opc, MRMSrcReg,
+              (outs _.KRC:$dst), (ins _.KRCWM:$mask, _.RC:$src1, _.RC:$src2),
+              !strconcat(OpcodeStr, "\t{$src2, $src1, $dst {${mask}}|",
+                          "$dst {${mask}}, $src1, $src2}"),
+              [(set _.KRC:$dst, (and _.KRCWM:$mask,
+                                   (OpNode (_.VT _.RC:$src1), (_.VT _.RC:$src2))))],
+              IIC_SSE_ALU_F32P_RR>, EVEX_4V, EVEX_K;
+  let mayLoad = 1 in
+  def rmk : AVX512BI<opc, MRMSrcMem,
+              (outs _.KRC:$dst), (ins _.KRCWM:$mask, _.RC:$src1, _.MemOp:$src2),
+              !strconcat(OpcodeStr, "\t{$src2, $src1, $dst {${mask}}|",
+                          "$dst {${mask}}, $src1, $src2}"),
+              [(set _.KRC:$dst, (and _.KRCWM:$mask,
+                                   (OpNode (_.VT _.RC:$src1),
+                                       (_.VT (bitconvert
+                                              (_.LdFrag addr:$src2))))))],
+              IIC_SSE_ALU_F32P_RM>, EVEX_4V, EVEX_K;
+}
+
+multiclass avx512_icmp_packed_rmb<bits<8> opc, string OpcodeStr, SDNode OpNode,
+              X86VectorVTInfo _> {
+  let mayLoad = 1 in {
+  def rmb : AVX512BI<opc, MRMSrcMem,
+              (outs _.KRC:$dst), (ins _.RC:$src1, _.ScalarMemOp:$src2),
+              !strconcat(OpcodeStr, "\t{${src2}", _.BroadcastStr, ", $src1, $dst",
+                                    "|$dst, $src1, ${src2}", _.BroadcastStr, "}"),
+              [(set _.KRC:$dst, (OpNode (_.VT _.RC:$src1),
+                              (X86VBroadcast (_.ScalarLdFrag addr:$src2))))],
+              IIC_SSE_ALU_F32P_RM>, EVEX_4V, EVEX_B;
+  def rmbk : AVX512BI<opc, MRMSrcMem,
+               (outs _.KRC:$dst), (ins _.KRCWM:$mask, _.RC:$src1,
+                                       _.ScalarMemOp:$src2),
+               !strconcat(OpcodeStr,
+                          "\t{${src2}", _.BroadcastStr, ", $src1, $dst {${mask}}|",
+                          "$dst {${mask}}, $src1, ${src2}", _.BroadcastStr, "}"),
+               [(set _.KRC:$dst, (and _.KRCWM:$mask,
+                                      (OpNode (_.VT _.RC:$src1),
+                                        (X86VBroadcast
+                                          (_.ScalarLdFrag addr:$src2)))))],
+               IIC_SSE_ALU_F32P_RM>, EVEX_4V, EVEX_K, EVEX_B;
+  }
 }
 
-defm VPCMPEQDZ : avx512_icmp_packed<0x76, "vpcmpeqd", VK16, VR512, i512mem, 
-                           memopv16i32, X86pcmpeqm, v16i32>, EVEX_V512,
-                           EVEX_CD8<32, CD8VF>;
-defm VPCMPEQQZ : avx512_icmp_packed<0x29, "vpcmpeqq", VK8, VR512, i512mem, 
-                           memopv8i64, X86pcmpeqm, v8i64>, T8PD, EVEX_V512,
-                           VEX_W, EVEX_CD8<64, CD8VF>;
+multiclass avx512_icmp_packed_vl<bits<8> opc, string OpcodeStr, SDNode OpNode,
+                                 AVX512VLVectorVTInfo VTInfo, Predicate prd> {
+  let Predicates = [prd] in
+  defm Z : avx512_icmp_packed<opc, OpcodeStr, OpNode, VTInfo.info512>,
+           EVEX_V512;
 
-defm VPCMPGTDZ : avx512_icmp_packed<0x66, "vpcmpgtd", VK16, VR512, i512mem, 
-                           memopv16i32, X86pcmpgtm, v16i32>, EVEX_V512,
-                           EVEX_CD8<32, CD8VF>;
-defm VPCMPGTQZ : avx512_icmp_packed<0x37, "vpcmpgtq", VK8, VR512, i512mem, 
-                           memopv8i64, X86pcmpgtm, v8i64>, T8PD, EVEX_V512,
-                           VEX_W, EVEX_CD8<64, CD8VF>;
+  let Predicates = [prd, HasVLX] in {
+    defm Z256 : avx512_icmp_packed<opc, OpcodeStr, OpNode, VTInfo.info256>,
+                EVEX_V256;
+    defm Z128 : avx512_icmp_packed<opc, OpcodeStr, OpNode, VTInfo.info128>,
+                EVEX_V128;
+  }
+}
+
+multiclass avx512_icmp_packed_rmb_vl<bits<8> opc, string OpcodeStr,
+                                  SDNode OpNode, AVX512VLVectorVTInfo VTInfo,
+                                  Predicate prd> {
+  let Predicates = [prd] in
+  defm Z : avx512_icmp_packed_rmb<opc, OpcodeStr, OpNode, VTInfo.info512>,
+           EVEX_V512;
+
+  let Predicates = [prd, HasVLX] in {
+    defm Z256 : avx512_icmp_packed_rmb<opc, OpcodeStr, OpNode, VTInfo.info256>,
+                EVEX_V256;
+    defm Z128 : avx512_icmp_packed_rmb<opc, OpcodeStr, OpNode, VTInfo.info128>,
+                EVEX_V128;
+  }
+}
+
+defm VPCMPEQB : avx512_icmp_packed_vl<0x74, "vpcmpeqb", X86pcmpeqm,
+                      avx512vl_i8_info, HasBWI>,
+                EVEX_CD8<8, CD8VF>;
+
+defm VPCMPEQW : avx512_icmp_packed_vl<0x75, "vpcmpeqw", X86pcmpeqm,
+                      avx512vl_i16_info, HasBWI>,
+                EVEX_CD8<16, CD8VF>;
+
+defm VPCMPEQD : avx512_icmp_packed_vl<0x76, "vpcmpeqd", X86pcmpeqm,
+                      avx512vl_i32_info, HasAVX512>,
+                avx512_icmp_packed_rmb_vl<0x76, "vpcmpeqd", X86pcmpeqm,
+                      avx512vl_i32_info, HasAVX512>,
+                EVEX_CD8<32, CD8VF>;
+
+defm VPCMPEQQ : avx512_icmp_packed_vl<0x29, "vpcmpeqq", X86pcmpeqm,
+                      avx512vl_i64_info, HasAVX512>,
+                avx512_icmp_packed_rmb_vl<0x29, "vpcmpeqq", X86pcmpeqm,
+                      avx512vl_i64_info, HasAVX512>,
+                T8PD, VEX_W, EVEX_CD8<64, CD8VF>;
+
+defm VPCMPGTB : avx512_icmp_packed_vl<0x64, "vpcmpgtb", X86pcmpgtm,
+                      avx512vl_i8_info, HasBWI>,
+                EVEX_CD8<8, CD8VF>;
+
+defm VPCMPGTW : avx512_icmp_packed_vl<0x65, "vpcmpgtw", X86pcmpgtm,
+                      avx512vl_i16_info, HasBWI>,
+                EVEX_CD8<16, CD8VF>;
+
+defm VPCMPGTD : avx512_icmp_packed_vl<0x66, "vpcmpgtd", X86pcmpgtm,
+                      avx512vl_i32_info, HasAVX512>,
+                avx512_icmp_packed_rmb_vl<0x66, "vpcmpgtd", X86pcmpgtm,
+                      avx512vl_i32_info, HasAVX512>,
+                EVEX_CD8<32, CD8VF>;
+
+defm VPCMPGTQ : avx512_icmp_packed_vl<0x37, "vpcmpgtq", X86pcmpgtm,
+                      avx512vl_i64_info, HasAVX512>,
+                avx512_icmp_packed_rmb_vl<0x37, "vpcmpgtq", X86pcmpgtm,
+                      avx512vl_i64_info, HasAVX512>,
+                T8PD, VEX_W, EVEX_CD8<64, CD8VF>;
 
 def : Pat<(v8i1 (X86pcmpgtm (v8i32 VR256X:$src1), (v8i32 VR256X:$src2))),
             (COPY_TO_REGCLASS (VPCMPGTDZrr 
index 8cd7aa4..c98f60c 100644 (file)
 // CHECK:  encoding: [0x62,0xe1,0xdd,0x58,0xdb,0x8a,0xf8,0xfb,0xff,0xff]
           vpandq -1032(%rdx){1to8}, %zmm4, %zmm17
 
+// CHECK: vpcmpeqd %zmm10, %zmm2, %k5
+// CHECK:  encoding: [0x62,0xd1,0x6d,0x48,0x76,0xea]
+          vpcmpeqd %zmm10, %zmm2, %k5
+
+// CHECK: vpcmpeqd %zmm10, %zmm2, %k5 {%k7}
+// CHECK:  encoding: [0x62,0xd1,0x6d,0x4f,0x76,0xea]
+          vpcmpeqd %zmm10, %zmm2, %k5 {%k7}
+
+// CHECK: vpcmpeqd (%rcx), %zmm2, %k5
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x48,0x76,0x29]
+          vpcmpeqd (%rcx), %zmm2, %k5
+
+// CHECK: vpcmpeqd 291(%rax,%r14,8), %zmm2, %k5
+// CHECK:  encoding: [0x62,0xb1,0x6d,0x48,0x76,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpeqd 291(%rax,%r14,8), %zmm2, %k5
+
+// CHECK: vpcmpeqd (%rcx){1to16}, %zmm2, %k5
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x58,0x76,0x29]
+          vpcmpeqd (%rcx){1to16}, %zmm2, %k5
+
+// CHECK: vpcmpeqd 8128(%rdx), %zmm2, %k5
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x48,0x76,0x6a,0x7f]
+          vpcmpeqd 8128(%rdx), %zmm2, %k5
+
+// CHECK: vpcmpeqd 8192(%rdx), %zmm2, %k5
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x48,0x76,0xaa,0x00,0x20,0x00,0x00]
+          vpcmpeqd 8192(%rdx), %zmm2, %k5
+
+// CHECK: vpcmpeqd -8192(%rdx), %zmm2, %k5
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x48,0x76,0x6a,0x80]
+          vpcmpeqd -8192(%rdx), %zmm2, %k5
+
+// CHECK: vpcmpeqd -8256(%rdx), %zmm2, %k5
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x48,0x76,0xaa,0xc0,0xdf,0xff,0xff]
+          vpcmpeqd -8256(%rdx), %zmm2, %k5
+
+// CHECK: vpcmpeqd 508(%rdx){1to16}, %zmm2, %k5
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x58,0x76,0x6a,0x7f]
+          vpcmpeqd 508(%rdx){1to16}, %zmm2, %k5
+
+// CHECK: vpcmpeqd 512(%rdx){1to16}, %zmm2, %k5
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x58,0x76,0xaa,0x00,0x02,0x00,0x00]
+          vpcmpeqd 512(%rdx){1to16}, %zmm2, %k5
+
+// CHECK: vpcmpeqd -512(%rdx){1to16}, %zmm2, %k5
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x58,0x76,0x6a,0x80]
+          vpcmpeqd -512(%rdx){1to16}, %zmm2, %k5
+
+// CHECK: vpcmpeqd -516(%rdx){1to16}, %zmm2, %k5
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x58,0x76,0xaa,0xfc,0xfd,0xff,0xff]
+          vpcmpeqd -516(%rdx){1to16}, %zmm2, %k5
+
+// CHECK: vpcmpeqq %zmm2, %zmm22, %k3
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x40,0x29,0xda]
+          vpcmpeqq %zmm2, %zmm22, %k3
+
+// CHECK: vpcmpeqq %zmm2, %zmm22, %k3 {%k6}
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x46,0x29,0xda]
+          vpcmpeqq %zmm2, %zmm22, %k3 {%k6}
+
+// CHECK: vpcmpeqq (%rcx), %zmm22, %k3
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x40,0x29,0x19]
+          vpcmpeqq (%rcx), %zmm22, %k3
+
+// CHECK: vpcmpeqq 291(%rax,%r14,8), %zmm22, %k3
+// CHECK:  encoding: [0x62,0xb2,0xcd,0x40,0x29,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpeqq 291(%rax,%r14,8), %zmm22, %k3
+
+// CHECK: vpcmpeqq (%rcx){1to8}, %zmm22, %k3
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x50,0x29,0x19]
+          vpcmpeqq (%rcx){1to8}, %zmm22, %k3
+
+// CHECK: vpcmpeqq 8128(%rdx), %zmm22, %k3
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x40,0x29,0x5a,0x7f]
+          vpcmpeqq 8128(%rdx), %zmm22, %k3
+
+// CHECK: vpcmpeqq 8192(%rdx), %zmm22, %k3
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x40,0x29,0x9a,0x00,0x20,0x00,0x00]
+          vpcmpeqq 8192(%rdx), %zmm22, %k3
+
+// CHECK: vpcmpeqq -8192(%rdx), %zmm22, %k3
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x40,0x29,0x5a,0x80]
+          vpcmpeqq -8192(%rdx), %zmm22, %k3
+
+// CHECK: vpcmpeqq -8256(%rdx), %zmm22, %k3
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x40,0x29,0x9a,0xc0,0xdf,0xff,0xff]
+          vpcmpeqq -8256(%rdx), %zmm22, %k3
+
+// CHECK: vpcmpeqq 1016(%rdx){1to8}, %zmm22, %k3
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x50,0x29,0x5a,0x7f]
+          vpcmpeqq 1016(%rdx){1to8}, %zmm22, %k3
+
+// CHECK: vpcmpeqq 1024(%rdx){1to8}, %zmm22, %k3
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x50,0x29,0x9a,0x00,0x04,0x00,0x00]
+          vpcmpeqq 1024(%rdx){1to8}, %zmm22, %k3
+
+// CHECK: vpcmpeqq -1024(%rdx){1to8}, %zmm22, %k3
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x50,0x29,0x5a,0x80]
+          vpcmpeqq -1024(%rdx){1to8}, %zmm22, %k3
+
+// CHECK: vpcmpeqq -1032(%rdx){1to8}, %zmm22, %k3
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x50,0x29,0x9a,0xf8,0xfb,0xff,0xff]
+          vpcmpeqq -1032(%rdx){1to8}, %zmm22, %k3
+
+// CHECK: vpcmpgtd %zmm8, %zmm21, %k5
+// CHECK:  encoding: [0x62,0xd1,0x55,0x40,0x66,0xe8]
+          vpcmpgtd %zmm8, %zmm21, %k5
+
+// CHECK: vpcmpgtd %zmm8, %zmm21, %k5 {%k5}
+// CHECK:  encoding: [0x62,0xd1,0x55,0x45,0x66,0xe8]
+          vpcmpgtd %zmm8, %zmm21, %k5 {%k5}
+
+// CHECK: vpcmpgtd (%rcx), %zmm21, %k5
+// CHECK:  encoding: [0x62,0xf1,0x55,0x40,0x66,0x29]
+          vpcmpgtd (%rcx), %zmm21, %k5
+
+// CHECK: vpcmpgtd 291(%rax,%r14,8), %zmm21, %k5
+// CHECK:  encoding: [0x62,0xb1,0x55,0x40,0x66,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpgtd 291(%rax,%r14,8), %zmm21, %k5
+
+// CHECK: vpcmpgtd (%rcx){1to16}, %zmm21, %k5
+// CHECK:  encoding: [0x62,0xf1,0x55,0x50,0x66,0x29]
+          vpcmpgtd (%rcx){1to16}, %zmm21, %k5
+
+// CHECK: vpcmpgtd 8128(%rdx), %zmm21, %k5
+// CHECK:  encoding: [0x62,0xf1,0x55,0x40,0x66,0x6a,0x7f]
+          vpcmpgtd 8128(%rdx), %zmm21, %k5
+
+// CHECK: vpcmpgtd 8192(%rdx), %zmm21, %k5
+// CHECK:  encoding: [0x62,0xf1,0x55,0x40,0x66,0xaa,0x00,0x20,0x00,0x00]
+          vpcmpgtd 8192(%rdx), %zmm21, %k5
+
+// CHECK: vpcmpgtd -8192(%rdx), %zmm21, %k5
+// CHECK:  encoding: [0x62,0xf1,0x55,0x40,0x66,0x6a,0x80]
+          vpcmpgtd -8192(%rdx), %zmm21, %k5
+
+// CHECK: vpcmpgtd -8256(%rdx), %zmm21, %k5
+// CHECK:  encoding: [0x62,0xf1,0x55,0x40,0x66,0xaa,0xc0,0xdf,0xff,0xff]
+          vpcmpgtd -8256(%rdx), %zmm21, %k5
+
+// CHECK: vpcmpgtd 508(%rdx){1to16}, %zmm21, %k5
+// CHECK:  encoding: [0x62,0xf1,0x55,0x50,0x66,0x6a,0x7f]
+          vpcmpgtd 508(%rdx){1to16}, %zmm21, %k5
+
+// CHECK: vpcmpgtd 512(%rdx){1to16}, %zmm21, %k5
+// CHECK:  encoding: [0x62,0xf1,0x55,0x50,0x66,0xaa,0x00,0x02,0x00,0x00]
+          vpcmpgtd 512(%rdx){1to16}, %zmm21, %k5
+
+// CHECK: vpcmpgtd -512(%rdx){1to16}, %zmm21, %k5
+// CHECK:  encoding: [0x62,0xf1,0x55,0x50,0x66,0x6a,0x80]
+          vpcmpgtd -512(%rdx){1to16}, %zmm21, %k5
+
+// CHECK: vpcmpgtd -516(%rdx){1to16}, %zmm21, %k5
+// CHECK:  encoding: [0x62,0xf1,0x55,0x50,0x66,0xaa,0xfc,0xfd,0xff,0xff]
+          vpcmpgtd -516(%rdx){1to16}, %zmm21, %k5
+
+// CHECK: vpcmpgtq %zmm17, %zmm20, %k2
+// CHECK:  encoding: [0x62,0xb2,0xdd,0x40,0x37,0xd1]
+          vpcmpgtq %zmm17, %zmm20, %k2
+
+// CHECK: vpcmpgtq %zmm17, %zmm20, %k2 {%k3}
+// CHECK:  encoding: [0x62,0xb2,0xdd,0x43,0x37,0xd1]
+          vpcmpgtq %zmm17, %zmm20, %k2 {%k3}
+
+// CHECK: vpcmpgtq (%rcx), %zmm20, %k2
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x40,0x37,0x11]
+          vpcmpgtq (%rcx), %zmm20, %k2
+
+// CHECK: vpcmpgtq 291(%rax,%r14,8), %zmm20, %k2
+// CHECK:  encoding: [0x62,0xb2,0xdd,0x40,0x37,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpgtq 291(%rax,%r14,8), %zmm20, %k2
+
+// CHECK: vpcmpgtq (%rcx){1to8}, %zmm20, %k2
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x50,0x37,0x11]
+          vpcmpgtq (%rcx){1to8}, %zmm20, %k2
+
+// CHECK: vpcmpgtq 8128(%rdx), %zmm20, %k2
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x40,0x37,0x52,0x7f]
+          vpcmpgtq 8128(%rdx), %zmm20, %k2
+
+// CHECK: vpcmpgtq 8192(%rdx), %zmm20, %k2
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x40,0x37,0x92,0x00,0x20,0x00,0x00]
+          vpcmpgtq 8192(%rdx), %zmm20, %k2
+
+// CHECK: vpcmpgtq -8192(%rdx), %zmm20, %k2
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x40,0x37,0x52,0x80]
+          vpcmpgtq -8192(%rdx), %zmm20, %k2
+
+// CHECK: vpcmpgtq -8256(%rdx), %zmm20, %k2
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x40,0x37,0x92,0xc0,0xdf,0xff,0xff]
+          vpcmpgtq -8256(%rdx), %zmm20, %k2
+
+// CHECK: vpcmpgtq 1016(%rdx){1to8}, %zmm20, %k2
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x50,0x37,0x52,0x7f]
+          vpcmpgtq 1016(%rdx){1to8}, %zmm20, %k2
+
+// CHECK: vpcmpgtq 1024(%rdx){1to8}, %zmm20, %k2
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x50,0x37,0x92,0x00,0x04,0x00,0x00]
+          vpcmpgtq 1024(%rdx){1to8}, %zmm20, %k2
+
+// CHECK: vpcmpgtq -1024(%rdx){1to8}, %zmm20, %k2
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x50,0x37,0x52,0x80]
+          vpcmpgtq -1024(%rdx){1to8}, %zmm20, %k2
+
+// CHECK: vpcmpgtq -1032(%rdx){1to8}, %zmm20, %k2
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x50,0x37,0x92,0xf8,0xfb,0xff,0xff]
+          vpcmpgtq -1032(%rdx){1to8}, %zmm20, %k2
+
 // CHECK: vpmaxsd %zmm16, %zmm8, %zmm6
 // CHECK:  encoding: [0x62,0xb2,0x3d,0x48,0x3d,0xf0]
           vpmaxsd %zmm16, %zmm8, %zmm6
index 9418a5b..ce688a1 100644 (file)
@@ -1,5 +1,133 @@
 // RUN: llvm-mc -triple x86_64-unknown-unknown -mcpu=skx  --show-encoding %s | FileCheck %s
 
+// CHECK: vpcmpeqb %zmm26, %zmm26, %k4
+// CHECK:  encoding: [0x62,0x91,0x2d,0x40,0x74,0xe2]
+          vpcmpeqb %zmm26, %zmm26, %k4
+
+// CHECK: vpcmpeqb %zmm26, %zmm26, %k4 {%k6}
+// CHECK:  encoding: [0x62,0x91,0x2d,0x46,0x74,0xe2]
+          vpcmpeqb %zmm26, %zmm26, %k4 {%k6}
+
+// CHECK: vpcmpeqb (%rcx), %zmm26, %k4
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x40,0x74,0x21]
+          vpcmpeqb (%rcx), %zmm26, %k4
+
+// CHECK: vpcmpeqb 291(%rax,%r14,8), %zmm26, %k4
+// CHECK:  encoding: [0x62,0xb1,0x2d,0x40,0x74,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpeqb 291(%rax,%r14,8), %zmm26, %k4
+
+// CHECK: vpcmpeqb 8128(%rdx), %zmm26, %k4
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x40,0x74,0x62,0x7f]
+          vpcmpeqb 8128(%rdx), %zmm26, %k4
+
+// CHECK: vpcmpeqb 8192(%rdx), %zmm26, %k4
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x40,0x74,0xa2,0x00,0x20,0x00,0x00]
+          vpcmpeqb 8192(%rdx), %zmm26, %k4
+
+// CHECK: vpcmpeqb -8192(%rdx), %zmm26, %k4
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x40,0x74,0x62,0x80]
+          vpcmpeqb -8192(%rdx), %zmm26, %k4
+
+// CHECK: vpcmpeqb -8256(%rdx), %zmm26, %k4
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x40,0x74,0xa2,0xc0,0xdf,0xff,0xff]
+          vpcmpeqb -8256(%rdx), %zmm26, %k4
+
+// CHECK: vpcmpeqw %zmm19, %zmm23, %k5
+// CHECK:  encoding: [0x62,0xb1,0x45,0x40,0x75,0xeb]
+          vpcmpeqw %zmm19, %zmm23, %k5
+
+// CHECK: vpcmpeqw %zmm19, %zmm23, %k5 {%k7}
+// CHECK:  encoding: [0x62,0xb1,0x45,0x47,0x75,0xeb]
+          vpcmpeqw %zmm19, %zmm23, %k5 {%k7}
+
+// CHECK: vpcmpeqw (%rcx), %zmm23, %k5
+// CHECK:  encoding: [0x62,0xf1,0x45,0x40,0x75,0x29]
+          vpcmpeqw (%rcx), %zmm23, %k5
+
+// CHECK: vpcmpeqw 291(%rax,%r14,8), %zmm23, %k5
+// CHECK:  encoding: [0x62,0xb1,0x45,0x40,0x75,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpeqw 291(%rax,%r14,8), %zmm23, %k5
+
+// CHECK: vpcmpeqw 8128(%rdx), %zmm23, %k5
+// CHECK:  encoding: [0x62,0xf1,0x45,0x40,0x75,0x6a,0x7f]
+          vpcmpeqw 8128(%rdx), %zmm23, %k5
+
+// CHECK: vpcmpeqw 8192(%rdx), %zmm23, %k5
+// CHECK:  encoding: [0x62,0xf1,0x45,0x40,0x75,0xaa,0x00,0x20,0x00,0x00]
+          vpcmpeqw 8192(%rdx), %zmm23, %k5
+
+// CHECK: vpcmpeqw -8192(%rdx), %zmm23, %k5
+// CHECK:  encoding: [0x62,0xf1,0x45,0x40,0x75,0x6a,0x80]
+          vpcmpeqw -8192(%rdx), %zmm23, %k5
+
+// CHECK: vpcmpeqw -8256(%rdx), %zmm23, %k5
+// CHECK:  encoding: [0x62,0xf1,0x45,0x40,0x75,0xaa,0xc0,0xdf,0xff,0xff]
+          vpcmpeqw -8256(%rdx), %zmm23, %k5
+
+// CHECK: vpcmpgtb %zmm20, %zmm30, %k4
+// CHECK:  encoding: [0x62,0xb1,0x0d,0x40,0x64,0xe4]
+          vpcmpgtb %zmm20, %zmm30, %k4
+
+// CHECK: vpcmpgtb %zmm20, %zmm30, %k4 {%k1}
+// CHECK:  encoding: [0x62,0xb1,0x0d,0x41,0x64,0xe4]
+          vpcmpgtb %zmm20, %zmm30, %k4 {%k1}
+
+// CHECK: vpcmpgtb (%rcx), %zmm30, %k4
+// CHECK:  encoding: [0x62,0xf1,0x0d,0x40,0x64,0x21]
+          vpcmpgtb (%rcx), %zmm30, %k4
+
+// CHECK: vpcmpgtb 291(%rax,%r14,8), %zmm30, %k4
+// CHECK:  encoding: [0x62,0xb1,0x0d,0x40,0x64,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpgtb 291(%rax,%r14,8), %zmm30, %k4
+
+// CHECK: vpcmpgtb 8128(%rdx), %zmm30, %k4
+// CHECK:  encoding: [0x62,0xf1,0x0d,0x40,0x64,0x62,0x7f]
+          vpcmpgtb 8128(%rdx), %zmm30, %k4
+
+// CHECK: vpcmpgtb 8192(%rdx), %zmm30, %k4
+// CHECK:  encoding: [0x62,0xf1,0x0d,0x40,0x64,0xa2,0x00,0x20,0x00,0x00]
+          vpcmpgtb 8192(%rdx), %zmm30, %k4
+
+// CHECK: vpcmpgtb -8192(%rdx), %zmm30, %k4
+// CHECK:  encoding: [0x62,0xf1,0x0d,0x40,0x64,0x62,0x80]
+          vpcmpgtb -8192(%rdx), %zmm30, %k4
+
+// CHECK: vpcmpgtb -8256(%rdx), %zmm30, %k4
+// CHECK:  encoding: [0x62,0xf1,0x0d,0x40,0x64,0xa2,0xc0,0xdf,0xff,0xff]
+          vpcmpgtb -8256(%rdx), %zmm30, %k4
+
+// CHECK: vpcmpgtw %zmm21, %zmm23, %k5
+// CHECK:  encoding: [0x62,0xb1,0x45,0x40,0x65,0xed]
+          vpcmpgtw %zmm21, %zmm23, %k5
+
+// CHECK: vpcmpgtw %zmm21, %zmm23, %k5 {%k7}
+// CHECK:  encoding: [0x62,0xb1,0x45,0x47,0x65,0xed]
+          vpcmpgtw %zmm21, %zmm23, %k5 {%k7}
+
+// CHECK: vpcmpgtw (%rcx), %zmm23, %k5
+// CHECK:  encoding: [0x62,0xf1,0x45,0x40,0x65,0x29]
+          vpcmpgtw (%rcx), %zmm23, %k5
+
+// CHECK: vpcmpgtw 291(%rax,%r14,8), %zmm23, %k5
+// CHECK:  encoding: [0x62,0xb1,0x45,0x40,0x65,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpgtw 291(%rax,%r14,8), %zmm23, %k5
+
+// CHECK: vpcmpgtw 8128(%rdx), %zmm23, %k5
+// CHECK:  encoding: [0x62,0xf1,0x45,0x40,0x65,0x6a,0x7f]
+          vpcmpgtw 8128(%rdx), %zmm23, %k5
+
+// CHECK: vpcmpgtw 8192(%rdx), %zmm23, %k5
+// CHECK:  encoding: [0x62,0xf1,0x45,0x40,0x65,0xaa,0x00,0x20,0x00,0x00]
+          vpcmpgtw 8192(%rdx), %zmm23, %k5
+
+// CHECK: vpcmpgtw -8192(%rdx), %zmm23, %k5
+// CHECK:  encoding: [0x62,0xf1,0x45,0x40,0x65,0x6a,0x80]
+          vpcmpgtw -8192(%rdx), %zmm23, %k5
+
+// CHECK: vpcmpgtw -8256(%rdx), %zmm23, %k5
+// CHECK:  encoding: [0x62,0xf1,0x45,0x40,0x65,0xaa,0xc0,0xdf,0xff,0xff]
+          vpcmpgtw -8256(%rdx), %zmm23, %k5
+
 // CHECK: vmovdqu8 %zmm19, %zmm29
 // CHECK:  encoding: [0x62,0x21,0x7f,0x48,0x6f,0xeb]
           vmovdqu8 %zmm19, %zmm29
index c8e2c19..f67d9f1 100644 (file)
@@ -1,5 +1,261 @@
 // RUN: llvm-mc -triple x86_64-unknown-unknown -mcpu=skx  --show-encoding %s | FileCheck %s
 
+// CHECK: vpcmpeqb %xmm21, %xmm21, %k4
+// CHECK:  encoding: [0x62,0xb1,0x55,0x00,0x74,0xe5]
+          vpcmpeqb %xmm21, %xmm21, %k4
+
+// CHECK: vpcmpeqb %xmm21, %xmm21, %k4 {%k3}
+// CHECK:  encoding: [0x62,0xb1,0x55,0x03,0x74,0xe5]
+          vpcmpeqb %xmm21, %xmm21, %k4 {%k3}
+
+// CHECK: vpcmpeqb (%rcx), %xmm21, %k4
+// CHECK:  encoding: [0x62,0xf1,0x55,0x00,0x74,0x21]
+          vpcmpeqb (%rcx), %xmm21, %k4
+
+// CHECK: vpcmpeqb 291(%rax,%r14,8), %xmm21, %k4
+// CHECK:  encoding: [0x62,0xb1,0x55,0x00,0x74,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpeqb 291(%rax,%r14,8), %xmm21, %k4
+
+// CHECK: vpcmpeqb 2032(%rdx), %xmm21, %k4
+// CHECK:  encoding: [0x62,0xf1,0x55,0x00,0x74,0x62,0x7f]
+          vpcmpeqb 2032(%rdx), %xmm21, %k4
+
+// CHECK: vpcmpeqb 2048(%rdx), %xmm21, %k4
+// CHECK:  encoding: [0x62,0xf1,0x55,0x00,0x74,0xa2,0x00,0x08,0x00,0x00]
+          vpcmpeqb 2048(%rdx), %xmm21, %k4
+
+// CHECK: vpcmpeqb -2048(%rdx), %xmm21, %k4
+// CHECK:  encoding: [0x62,0xf1,0x55,0x00,0x74,0x62,0x80]
+          vpcmpeqb -2048(%rdx), %xmm21, %k4
+
+// CHECK: vpcmpeqb -2064(%rdx), %xmm21, %k4
+// CHECK:  encoding: [0x62,0xf1,0x55,0x00,0x74,0xa2,0xf0,0xf7,0xff,0xff]
+          vpcmpeqb -2064(%rdx), %xmm21, %k4
+
+// CHECK: vpcmpeqb %ymm18, %ymm21, %k4
+// CHECK:  encoding: [0x62,0xb1,0x55,0x20,0x74,0xe2]
+          vpcmpeqb %ymm18, %ymm21, %k4
+
+// CHECK: vpcmpeqb %ymm18, %ymm21, %k4 {%k1}
+// CHECK:  encoding: [0x62,0xb1,0x55,0x21,0x74,0xe2]
+          vpcmpeqb %ymm18, %ymm21, %k4 {%k1}
+
+// CHECK: vpcmpeqb (%rcx), %ymm21, %k4
+// CHECK:  encoding: [0x62,0xf1,0x55,0x20,0x74,0x21]
+          vpcmpeqb (%rcx), %ymm21, %k4
+
+// CHECK: vpcmpeqb 291(%rax,%r14,8), %ymm21, %k4
+// CHECK:  encoding: [0x62,0xb1,0x55,0x20,0x74,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpeqb 291(%rax,%r14,8), %ymm21, %k4
+
+// CHECK: vpcmpeqb 4064(%rdx), %ymm21, %k4
+// CHECK:  encoding: [0x62,0xf1,0x55,0x20,0x74,0x62,0x7f]
+          vpcmpeqb 4064(%rdx), %ymm21, %k4
+
+// CHECK: vpcmpeqb 4096(%rdx), %ymm21, %k4
+// CHECK:  encoding: [0x62,0xf1,0x55,0x20,0x74,0xa2,0x00,0x10,0x00,0x00]
+          vpcmpeqb 4096(%rdx), %ymm21, %k4
+
+// CHECK: vpcmpeqb -4096(%rdx), %ymm21, %k4
+// CHECK:  encoding: [0x62,0xf1,0x55,0x20,0x74,0x62,0x80]
+          vpcmpeqb -4096(%rdx), %ymm21, %k4
+
+// CHECK: vpcmpeqb -4128(%rdx), %ymm21, %k4
+// CHECK:  encoding: [0x62,0xf1,0x55,0x20,0x74,0xa2,0xe0,0xef,0xff,0xff]
+          vpcmpeqb -4128(%rdx), %ymm21, %k4
+
+// CHECK: vpcmpeqw %xmm27, %xmm30, %k3
+// CHECK:  encoding: [0x62,0x91,0x0d,0x00,0x75,0xdb]
+          vpcmpeqw %xmm27, %xmm30, %k3
+
+// CHECK: vpcmpeqw %xmm27, %xmm30, %k3 {%k1}
+// CHECK:  encoding: [0x62,0x91,0x0d,0x01,0x75,0xdb]
+          vpcmpeqw %xmm27, %xmm30, %k3 {%k1}
+
+// CHECK: vpcmpeqw (%rcx), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf1,0x0d,0x00,0x75,0x19]
+          vpcmpeqw (%rcx), %xmm30, %k3
+
+// CHECK: vpcmpeqw 291(%rax,%r14,8), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xb1,0x0d,0x00,0x75,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpeqw 291(%rax,%r14,8), %xmm30, %k3
+
+// CHECK: vpcmpeqw 2032(%rdx), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf1,0x0d,0x00,0x75,0x5a,0x7f]
+          vpcmpeqw 2032(%rdx), %xmm30, %k3
+
+// CHECK: vpcmpeqw 2048(%rdx), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf1,0x0d,0x00,0x75,0x9a,0x00,0x08,0x00,0x00]
+          vpcmpeqw 2048(%rdx), %xmm30, %k3
+
+// CHECK: vpcmpeqw -2048(%rdx), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf1,0x0d,0x00,0x75,0x5a,0x80]
+          vpcmpeqw -2048(%rdx), %xmm30, %k3
+
+// CHECK: vpcmpeqw -2064(%rdx), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf1,0x0d,0x00,0x75,0x9a,0xf0,0xf7,0xff,0xff]
+          vpcmpeqw -2064(%rdx), %xmm30, %k3
+
+// CHECK: vpcmpeqw %ymm29, %ymm20, %k2
+// CHECK:  encoding: [0x62,0x91,0x5d,0x20,0x75,0xd5]
+          vpcmpeqw %ymm29, %ymm20, %k2
+
+// CHECK: vpcmpeqw %ymm29, %ymm20, %k2 {%k5}
+// CHECK:  encoding: [0x62,0x91,0x5d,0x25,0x75,0xd5]
+          vpcmpeqw %ymm29, %ymm20, %k2 {%k5}
+
+// CHECK: vpcmpeqw (%rcx), %ymm20, %k2
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x20,0x75,0x11]
+          vpcmpeqw (%rcx), %ymm20, %k2
+
+// CHECK: vpcmpeqw 291(%rax,%r14,8), %ymm20, %k2
+// CHECK:  encoding: [0x62,0xb1,0x5d,0x20,0x75,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpeqw 291(%rax,%r14,8), %ymm20, %k2
+
+// CHECK: vpcmpeqw 4064(%rdx), %ymm20, %k2
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x20,0x75,0x52,0x7f]
+          vpcmpeqw 4064(%rdx), %ymm20, %k2
+
+// CHECK: vpcmpeqw 4096(%rdx), %ymm20, %k2
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x20,0x75,0x92,0x00,0x10,0x00,0x00]
+          vpcmpeqw 4096(%rdx), %ymm20, %k2
+
+// CHECK: vpcmpeqw -4096(%rdx), %ymm20, %k2
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x20,0x75,0x52,0x80]
+          vpcmpeqw -4096(%rdx), %ymm20, %k2
+
+// CHECK: vpcmpeqw -4128(%rdx), %ymm20, %k2
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x20,0x75,0x92,0xe0,0xef,0xff,0xff]
+          vpcmpeqw -4128(%rdx), %ymm20, %k2
+
+// CHECK: vpcmpgtb %xmm17, %xmm30, %k3
+// CHECK:  encoding: [0x62,0xb1,0x0d,0x00,0x64,0xd9]
+          vpcmpgtb %xmm17, %xmm30, %k3
+
+// CHECK: vpcmpgtb %xmm17, %xmm30, %k3 {%k7}
+// CHECK:  encoding: [0x62,0xb1,0x0d,0x07,0x64,0xd9]
+          vpcmpgtb %xmm17, %xmm30, %k3 {%k7}
+
+// CHECK: vpcmpgtb (%rcx), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf1,0x0d,0x00,0x64,0x19]
+          vpcmpgtb (%rcx), %xmm30, %k3
+
+// CHECK: vpcmpgtb 291(%rax,%r14,8), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xb1,0x0d,0x00,0x64,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpgtb 291(%rax,%r14,8), %xmm30, %k3
+
+// CHECK: vpcmpgtb 2032(%rdx), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf1,0x0d,0x00,0x64,0x5a,0x7f]
+          vpcmpgtb 2032(%rdx), %xmm30, %k3
+
+// CHECK: vpcmpgtb 2048(%rdx), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf1,0x0d,0x00,0x64,0x9a,0x00,0x08,0x00,0x00]
+          vpcmpgtb 2048(%rdx), %xmm30, %k3
+
+// CHECK: vpcmpgtb -2048(%rdx), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf1,0x0d,0x00,0x64,0x5a,0x80]
+          vpcmpgtb -2048(%rdx), %xmm30, %k3
+
+// CHECK: vpcmpgtb -2064(%rdx), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf1,0x0d,0x00,0x64,0x9a,0xf0,0xf7,0xff,0xff]
+          vpcmpgtb -2064(%rdx), %xmm30, %k3
+
+// CHECK: vpcmpgtb %ymm17, %ymm17, %k2
+// CHECK:  encoding: [0x62,0xb1,0x75,0x20,0x64,0xd1]
+          vpcmpgtb %ymm17, %ymm17, %k2
+
+// CHECK: vpcmpgtb %ymm17, %ymm17, %k2 {%k4}
+// CHECK:  encoding: [0x62,0xb1,0x75,0x24,0x64,0xd1]
+          vpcmpgtb %ymm17, %ymm17, %k2 {%k4}
+
+// CHECK: vpcmpgtb (%rcx), %ymm17, %k2
+// CHECK:  encoding: [0x62,0xf1,0x75,0x20,0x64,0x11]
+          vpcmpgtb (%rcx), %ymm17, %k2
+
+// CHECK: vpcmpgtb 291(%rax,%r14,8), %ymm17, %k2
+// CHECK:  encoding: [0x62,0xb1,0x75,0x20,0x64,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpgtb 291(%rax,%r14,8), %ymm17, %k2
+
+// CHECK: vpcmpgtb 4064(%rdx), %ymm17, %k2
+// CHECK:  encoding: [0x62,0xf1,0x75,0x20,0x64,0x52,0x7f]
+          vpcmpgtb 4064(%rdx), %ymm17, %k2
+
+// CHECK: vpcmpgtb 4096(%rdx), %ymm17, %k2
+// CHECK:  encoding: [0x62,0xf1,0x75,0x20,0x64,0x92,0x00,0x10,0x00,0x00]
+          vpcmpgtb 4096(%rdx), %ymm17, %k2
+
+// CHECK: vpcmpgtb -4096(%rdx), %ymm17, %k2
+// CHECK:  encoding: [0x62,0xf1,0x75,0x20,0x64,0x52,0x80]
+          vpcmpgtb -4096(%rdx), %ymm17, %k2
+
+// CHECK: vpcmpgtb -4128(%rdx), %ymm17, %k2
+// CHECK:  encoding: [0x62,0xf1,0x75,0x20,0x64,0x92,0xe0,0xef,0xff,0xff]
+          vpcmpgtb -4128(%rdx), %ymm17, %k2
+
+// CHECK: vpcmpgtw %xmm22, %xmm28, %k2
+// CHECK:  encoding: [0x62,0xb1,0x1d,0x00,0x65,0xd6]
+          vpcmpgtw %xmm22, %xmm28, %k2
+
+// CHECK: vpcmpgtw %xmm22, %xmm28, %k2 {%k7}
+// CHECK:  encoding: [0x62,0xb1,0x1d,0x07,0x65,0xd6]
+          vpcmpgtw %xmm22, %xmm28, %k2 {%k7}
+
+// CHECK: vpcmpgtw (%rcx), %xmm28, %k2
+// CHECK:  encoding: [0x62,0xf1,0x1d,0x00,0x65,0x11]
+          vpcmpgtw (%rcx), %xmm28, %k2
+
+// CHECK: vpcmpgtw 291(%rax,%r14,8), %xmm28, %k2
+// CHECK:  encoding: [0x62,0xb1,0x1d,0x00,0x65,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpgtw 291(%rax,%r14,8), %xmm28, %k2
+
+// CHECK: vpcmpgtw 2032(%rdx), %xmm28, %k2
+// CHECK:  encoding: [0x62,0xf1,0x1d,0x00,0x65,0x52,0x7f]
+          vpcmpgtw 2032(%rdx), %xmm28, %k2
+
+// CHECK: vpcmpgtw 2048(%rdx), %xmm28, %k2
+// CHECK:  encoding: [0x62,0xf1,0x1d,0x00,0x65,0x92,0x00,0x08,0x00,0x00]
+          vpcmpgtw 2048(%rdx), %xmm28, %k2
+
+// CHECK: vpcmpgtw -2048(%rdx), %xmm28, %k2
+// CHECK:  encoding: [0x62,0xf1,0x1d,0x00,0x65,0x52,0x80]
+          vpcmpgtw -2048(%rdx), %xmm28, %k2
+
+// CHECK: vpcmpgtw -2064(%rdx), %xmm28, %k2
+// CHECK:  encoding: [0x62,0xf1,0x1d,0x00,0x65,0x92,0xf0,0xf7,0xff,0xff]
+          vpcmpgtw -2064(%rdx), %xmm28, %k2
+
+// CHECK: vpcmpgtw %ymm26, %ymm20, %k5
+// CHECK:  encoding: [0x62,0x91,0x5d,0x20,0x65,0xea]
+          vpcmpgtw %ymm26, %ymm20, %k5
+
+// CHECK: vpcmpgtw %ymm26, %ymm20, %k5 {%k2}
+// CHECK:  encoding: [0x62,0x91,0x5d,0x22,0x65,0xea]
+          vpcmpgtw %ymm26, %ymm20, %k5 {%k2}
+
+// CHECK: vpcmpgtw (%rcx), %ymm20, %k5
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x20,0x65,0x29]
+          vpcmpgtw (%rcx), %ymm20, %k5
+
+// CHECK: vpcmpgtw 291(%rax,%r14,8), %ymm20, %k5
+// CHECK:  encoding: [0x62,0xb1,0x5d,0x20,0x65,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpgtw 291(%rax,%r14,8), %ymm20, %k5
+
+// CHECK: vpcmpgtw 4064(%rdx), %ymm20, %k5
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x20,0x65,0x6a,0x7f]
+          vpcmpgtw 4064(%rdx), %ymm20, %k5
+
+// CHECK: vpcmpgtw 4096(%rdx), %ymm20, %k5
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x20,0x65,0xaa,0x00,0x10,0x00,0x00]
+          vpcmpgtw 4096(%rdx), %ymm20, %k5
+
+// CHECK: vpcmpgtw -4096(%rdx), %ymm20, %k5
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x20,0x65,0x6a,0x80]
+          vpcmpgtw -4096(%rdx), %ymm20, %k5
+
+// CHECK: vpcmpgtw -4128(%rdx), %ymm20, %k5
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x20,0x65,0xaa,0xe0,0xef,0xff,0xff]
+          vpcmpgtw -4128(%rdx), %ymm20, %k5
+
 // CHECK: vmovdqu8 %xmm23, %xmm26
 // CHECK:  encoding: [0x62,0x21,0x7f,0x08,0x6f,0xd7]
           vmovdqu8 %xmm23, %xmm26
index b529c03..2da03f4 100644 (file)
 // CHECK:  encoding: [0x62,0xe1,0x7c,0x28,0x10,0xaa,0xe0,0xef,0xff,0xff]
           vmovups -4128(%rdx), %ymm21
 
+// CHECK: vpcmpeqd %xmm24, %xmm29, %k3
+// CHECK:  encoding: [0x62,0x91,0x15,0x00,0x76,0xd8]
+          vpcmpeqd %xmm24, %xmm29, %k3
+
+// CHECK: vpcmpeqd %xmm24, %xmm29, %k3 {%k5}
+// CHECK:  encoding: [0x62,0x91,0x15,0x05,0x76,0xd8]
+          vpcmpeqd %xmm24, %xmm29, %k3 {%k5}
+
+// CHECK: vpcmpeqd (%rcx), %xmm29, %k3
+// CHECK:  encoding: [0x62,0xf1,0x15,0x00,0x76,0x19]
+          vpcmpeqd (%rcx), %xmm29, %k3
+
+// CHECK: vpcmpeqd 291(%rax,%r14,8), %xmm29, %k3
+// CHECK:  encoding: [0x62,0xb1,0x15,0x00,0x76,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpeqd 291(%rax,%r14,8), %xmm29, %k3
+
+// CHECK: vpcmpeqd (%rcx){1to4}, %xmm29, %k3
+// CHECK:  encoding: [0x62,0xf1,0x15,0x10,0x76,0x19]
+          vpcmpeqd (%rcx){1to4}, %xmm29, %k3
+
+// CHECK: vpcmpeqd 2032(%rdx), %xmm29, %k3
+// CHECK:  encoding: [0x62,0xf1,0x15,0x00,0x76,0x5a,0x7f]
+          vpcmpeqd 2032(%rdx), %xmm29, %k3
+
+// CHECK: vpcmpeqd 2048(%rdx), %xmm29, %k3
+// CHECK:  encoding: [0x62,0xf1,0x15,0x00,0x76,0x9a,0x00,0x08,0x00,0x00]
+          vpcmpeqd 2048(%rdx), %xmm29, %k3
+
+// CHECK: vpcmpeqd -2048(%rdx), %xmm29, %k3
+// CHECK:  encoding: [0x62,0xf1,0x15,0x00,0x76,0x5a,0x80]
+          vpcmpeqd -2048(%rdx), %xmm29, %k3
+
+// CHECK: vpcmpeqd -2064(%rdx), %xmm29, %k3
+// CHECK:  encoding: [0x62,0xf1,0x15,0x00,0x76,0x9a,0xf0,0xf7,0xff,0xff]
+          vpcmpeqd -2064(%rdx), %xmm29, %k3
+
+// CHECK: vpcmpeqd 508(%rdx){1to4}, %xmm29, %k3
+// CHECK:  encoding: [0x62,0xf1,0x15,0x10,0x76,0x5a,0x7f]
+          vpcmpeqd 508(%rdx){1to4}, %xmm29, %k3
+
+// CHECK: vpcmpeqd 512(%rdx){1to4}, %xmm29, %k3
+// CHECK:  encoding: [0x62,0xf1,0x15,0x10,0x76,0x9a,0x00,0x02,0x00,0x00]
+          vpcmpeqd 512(%rdx){1to4}, %xmm29, %k3
+
+// CHECK: vpcmpeqd -512(%rdx){1to4}, %xmm29, %k3
+// CHECK:  encoding: [0x62,0xf1,0x15,0x10,0x76,0x5a,0x80]
+          vpcmpeqd -512(%rdx){1to4}, %xmm29, %k3
+
+// CHECK: vpcmpeqd -516(%rdx){1to4}, %xmm29, %k3
+// CHECK:  encoding: [0x62,0xf1,0x15,0x10,0x76,0x9a,0xfc,0xfd,0xff,0xff]
+          vpcmpeqd -516(%rdx){1to4}, %xmm29, %k3
+
+// CHECK: vpcmpeqd %ymm20, %ymm26, %k5
+// CHECK:  encoding: [0x62,0xb1,0x2d,0x20,0x76,0xec]
+          vpcmpeqd %ymm20, %ymm26, %k5
+
+// CHECK: vpcmpeqd %ymm20, %ymm26, %k5 {%k5}
+// CHECK:  encoding: [0x62,0xb1,0x2d,0x25,0x76,0xec]
+          vpcmpeqd %ymm20, %ymm26, %k5 {%k5}
+
+// CHECK: vpcmpeqd (%rcx), %ymm26, %k5
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x20,0x76,0x29]
+          vpcmpeqd (%rcx), %ymm26, %k5
+
+// CHECK: vpcmpeqd 291(%rax,%r14,8), %ymm26, %k5
+// CHECK:  encoding: [0x62,0xb1,0x2d,0x20,0x76,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpeqd 291(%rax,%r14,8), %ymm26, %k5
+
+// CHECK: vpcmpeqd (%rcx){1to8}, %ymm26, %k5
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x30,0x76,0x29]
+          vpcmpeqd (%rcx){1to8}, %ymm26, %k5
+
+// CHECK: vpcmpeqd 4064(%rdx), %ymm26, %k5
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x20,0x76,0x6a,0x7f]
+          vpcmpeqd 4064(%rdx), %ymm26, %k5
+
+// CHECK: vpcmpeqd 4096(%rdx), %ymm26, %k5
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x20,0x76,0xaa,0x00,0x10,0x00,0x00]
+          vpcmpeqd 4096(%rdx), %ymm26, %k5
+
+// CHECK: vpcmpeqd -4096(%rdx), %ymm26, %k5
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x20,0x76,0x6a,0x80]
+          vpcmpeqd -4096(%rdx), %ymm26, %k5
+
+// CHECK: vpcmpeqd -4128(%rdx), %ymm26, %k5
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x20,0x76,0xaa,0xe0,0xef,0xff,0xff]
+          vpcmpeqd -4128(%rdx), %ymm26, %k5
+
+// CHECK: vpcmpeqd 508(%rdx){1to8}, %ymm26, %k5
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x30,0x76,0x6a,0x7f]
+          vpcmpeqd 508(%rdx){1to8}, %ymm26, %k5
+
+// CHECK: vpcmpeqd 512(%rdx){1to8}, %ymm26, %k5
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x30,0x76,0xaa,0x00,0x02,0x00,0x00]
+          vpcmpeqd 512(%rdx){1to8}, %ymm26, %k5
+
+// CHECK: vpcmpeqd -512(%rdx){1to8}, %ymm26, %k5
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x30,0x76,0x6a,0x80]
+          vpcmpeqd -512(%rdx){1to8}, %ymm26, %k5
+
+// CHECK: vpcmpeqd -516(%rdx){1to8}, %ymm26, %k5
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x30,0x76,0xaa,0xfc,0xfd,0xff,0xff]
+          vpcmpeqd -516(%rdx){1to8}, %ymm26, %k5
+
+// CHECK: vpcmpeqq %xmm29, %xmm20, %k3
+// CHECK:  encoding: [0x62,0x92,0xdd,0x00,0x29,0xdd]
+          vpcmpeqq %xmm29, %xmm20, %k3
+
+// CHECK: vpcmpeqq %xmm29, %xmm20, %k3 {%k3}
+// CHECK:  encoding: [0x62,0x92,0xdd,0x03,0x29,0xdd]
+          vpcmpeqq %xmm29, %xmm20, %k3 {%k3}
+
+// CHECK: vpcmpeqq (%rcx), %xmm20, %k3
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x00,0x29,0x19]
+          vpcmpeqq (%rcx), %xmm20, %k3
+
+// CHECK: vpcmpeqq 291(%rax,%r14,8), %xmm20, %k3
+// CHECK:  encoding: [0x62,0xb2,0xdd,0x00,0x29,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpeqq 291(%rax,%r14,8), %xmm20, %k3
+
+// CHECK: vpcmpeqq (%rcx){1to2}, %xmm20, %k3
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x10,0x29,0x19]
+          vpcmpeqq (%rcx){1to2}, %xmm20, %k3
+
+// CHECK: vpcmpeqq 2032(%rdx), %xmm20, %k3
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x00,0x29,0x5a,0x7f]
+          vpcmpeqq 2032(%rdx), %xmm20, %k3
+
+// CHECK: vpcmpeqq 2048(%rdx), %xmm20, %k3
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x00,0x29,0x9a,0x00,0x08,0x00,0x00]
+          vpcmpeqq 2048(%rdx), %xmm20, %k3
+
+// CHECK: vpcmpeqq -2048(%rdx), %xmm20, %k3
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x00,0x29,0x5a,0x80]
+          vpcmpeqq -2048(%rdx), %xmm20, %k3
+
+// CHECK: vpcmpeqq -2064(%rdx), %xmm20, %k3
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x00,0x29,0x9a,0xf0,0xf7,0xff,0xff]
+          vpcmpeqq -2064(%rdx), %xmm20, %k3
+
+// CHECK: vpcmpeqq 1016(%rdx){1to2}, %xmm20, %k3
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x10,0x29,0x5a,0x7f]
+          vpcmpeqq 1016(%rdx){1to2}, %xmm20, %k3
+
+// CHECK: vpcmpeqq 1024(%rdx){1to2}, %xmm20, %k3
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x10,0x29,0x9a,0x00,0x04,0x00,0x00]
+          vpcmpeqq 1024(%rdx){1to2}, %xmm20, %k3
+
+// CHECK: vpcmpeqq -1024(%rdx){1to2}, %xmm20, %k3
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x10,0x29,0x5a,0x80]
+          vpcmpeqq -1024(%rdx){1to2}, %xmm20, %k3
+
+// CHECK: vpcmpeqq -1032(%rdx){1to2}, %xmm20, %k3
+// CHECK:  encoding: [0x62,0xf2,0xdd,0x10,0x29,0x9a,0xf8,0xfb,0xff,0xff]
+          vpcmpeqq -1032(%rdx){1to2}, %xmm20, %k3
+
+// CHECK: vpcmpeqq %ymm23, %ymm30, %k5
+// CHECK:  encoding: [0x62,0xb2,0x8d,0x20,0x29,0xef]
+          vpcmpeqq %ymm23, %ymm30, %k5
+
+// CHECK: vpcmpeqq %ymm23, %ymm30, %k5 {%k6}
+// CHECK:  encoding: [0x62,0xb2,0x8d,0x26,0x29,0xef]
+          vpcmpeqq %ymm23, %ymm30, %k5 {%k6}
+
+// CHECK: vpcmpeqq (%rcx), %ymm30, %k5
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x20,0x29,0x29]
+          vpcmpeqq (%rcx), %ymm30, %k5
+
+// CHECK: vpcmpeqq 291(%rax,%r14,8), %ymm30, %k5
+// CHECK:  encoding: [0x62,0xb2,0x8d,0x20,0x29,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpeqq 291(%rax,%r14,8), %ymm30, %k5
+
+// CHECK: vpcmpeqq (%rcx){1to4}, %ymm30, %k5
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x30,0x29,0x29]
+          vpcmpeqq (%rcx){1to4}, %ymm30, %k5
+
+// CHECK: vpcmpeqq 4064(%rdx), %ymm30, %k5
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x20,0x29,0x6a,0x7f]
+          vpcmpeqq 4064(%rdx), %ymm30, %k5
+
+// CHECK: vpcmpeqq 4096(%rdx), %ymm30, %k5
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x20,0x29,0xaa,0x00,0x10,0x00,0x00]
+          vpcmpeqq 4096(%rdx), %ymm30, %k5
+
+// CHECK: vpcmpeqq -4096(%rdx), %ymm30, %k5
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x20,0x29,0x6a,0x80]
+          vpcmpeqq -4096(%rdx), %ymm30, %k5
+
+// CHECK: vpcmpeqq -4128(%rdx), %ymm30, %k5
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x20,0x29,0xaa,0xe0,0xef,0xff,0xff]
+          vpcmpeqq -4128(%rdx), %ymm30, %k5
+
+// CHECK: vpcmpeqq 1016(%rdx){1to4}, %ymm30, %k5
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x30,0x29,0x6a,0x7f]
+          vpcmpeqq 1016(%rdx){1to4}, %ymm30, %k5
+
+// CHECK: vpcmpeqq 1024(%rdx){1to4}, %ymm30, %k5
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x30,0x29,0xaa,0x00,0x04,0x00,0x00]
+          vpcmpeqq 1024(%rdx){1to4}, %ymm30, %k5
+
+// CHECK: vpcmpeqq -1024(%rdx){1to4}, %ymm30, %k5
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x30,0x29,0x6a,0x80]
+          vpcmpeqq -1024(%rdx){1to4}, %ymm30, %k5
+
+// CHECK: vpcmpeqq -1032(%rdx){1to4}, %ymm30, %k5
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x30,0x29,0xaa,0xf8,0xfb,0xff,0xff]
+          vpcmpeqq -1032(%rdx){1to4}, %ymm30, %k5
+
+// CHECK: vpcmpgtd %xmm20, %xmm29, %k4
+// CHECK:  encoding: [0x62,0xb1,0x15,0x00,0x66,0xe4]
+          vpcmpgtd %xmm20, %xmm29, %k4
+
+// CHECK: vpcmpgtd %xmm20, %xmm29, %k4 {%k2}
+// CHECK:  encoding: [0x62,0xb1,0x15,0x02,0x66,0xe4]
+          vpcmpgtd %xmm20, %xmm29, %k4 {%k2}
+
+// CHECK: vpcmpgtd (%rcx), %xmm29, %k4
+// CHECK:  encoding: [0x62,0xf1,0x15,0x00,0x66,0x21]
+          vpcmpgtd (%rcx), %xmm29, %k4
+
+// CHECK: vpcmpgtd 291(%rax,%r14,8), %xmm29, %k4
+// CHECK:  encoding: [0x62,0xb1,0x15,0x00,0x66,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpgtd 291(%rax,%r14,8), %xmm29, %k4
+
+// CHECK: vpcmpgtd (%rcx){1to4}, %xmm29, %k4
+// CHECK:  encoding: [0x62,0xf1,0x15,0x10,0x66,0x21]
+          vpcmpgtd (%rcx){1to4}, %xmm29, %k4
+
+// CHECK: vpcmpgtd 2032(%rdx), %xmm29, %k4
+// CHECK:  encoding: [0x62,0xf1,0x15,0x00,0x66,0x62,0x7f]
+          vpcmpgtd 2032(%rdx), %xmm29, %k4
+
+// CHECK: vpcmpgtd 2048(%rdx), %xmm29, %k4
+// CHECK:  encoding: [0x62,0xf1,0x15,0x00,0x66,0xa2,0x00,0x08,0x00,0x00]
+          vpcmpgtd 2048(%rdx), %xmm29, %k4
+
+// CHECK: vpcmpgtd -2048(%rdx), %xmm29, %k4
+// CHECK:  encoding: [0x62,0xf1,0x15,0x00,0x66,0x62,0x80]
+          vpcmpgtd -2048(%rdx), %xmm29, %k4
+
+// CHECK: vpcmpgtd -2064(%rdx), %xmm29, %k4
+// CHECK:  encoding: [0x62,0xf1,0x15,0x00,0x66,0xa2,0xf0,0xf7,0xff,0xff]
+          vpcmpgtd -2064(%rdx), %xmm29, %k4
+
+// CHECK: vpcmpgtd 508(%rdx){1to4}, %xmm29, %k4
+// CHECK:  encoding: [0x62,0xf1,0x15,0x10,0x66,0x62,0x7f]
+          vpcmpgtd 508(%rdx){1to4}, %xmm29, %k4
+
+// CHECK: vpcmpgtd 512(%rdx){1to4}, %xmm29, %k4
+// CHECK:  encoding: [0x62,0xf1,0x15,0x10,0x66,0xa2,0x00,0x02,0x00,0x00]
+          vpcmpgtd 512(%rdx){1to4}, %xmm29, %k4
+
+// CHECK: vpcmpgtd -512(%rdx){1to4}, %xmm29, %k4
+// CHECK:  encoding: [0x62,0xf1,0x15,0x10,0x66,0x62,0x80]
+          vpcmpgtd -512(%rdx){1to4}, %xmm29, %k4
+
+// CHECK: vpcmpgtd -516(%rdx){1to4}, %xmm29, %k4
+// CHECK:  encoding: [0x62,0xf1,0x15,0x10,0x66,0xa2,0xfc,0xfd,0xff,0xff]
+          vpcmpgtd -516(%rdx){1to4}, %xmm29, %k4
+
+// CHECK: vpcmpgtd %ymm17, %ymm22, %k2
+// CHECK:  encoding: [0x62,0xb1,0x4d,0x20,0x66,0xd1]
+          vpcmpgtd %ymm17, %ymm22, %k2
+
+// CHECK: vpcmpgtd %ymm17, %ymm22, %k2 {%k1}
+// CHECK:  encoding: [0x62,0xb1,0x4d,0x21,0x66,0xd1]
+          vpcmpgtd %ymm17, %ymm22, %k2 {%k1}
+
+// CHECK: vpcmpgtd (%rcx), %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf1,0x4d,0x20,0x66,0x11]
+          vpcmpgtd (%rcx), %ymm22, %k2
+
+// CHECK: vpcmpgtd 291(%rax,%r14,8), %ymm22, %k2
+// CHECK:  encoding: [0x62,0xb1,0x4d,0x20,0x66,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpgtd 291(%rax,%r14,8), %ymm22, %k2
+
+// CHECK: vpcmpgtd (%rcx){1to8}, %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf1,0x4d,0x30,0x66,0x11]
+          vpcmpgtd (%rcx){1to8}, %ymm22, %k2
+
+// CHECK: vpcmpgtd 4064(%rdx), %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf1,0x4d,0x20,0x66,0x52,0x7f]
+          vpcmpgtd 4064(%rdx), %ymm22, %k2
+
+// CHECK: vpcmpgtd 4096(%rdx), %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf1,0x4d,0x20,0x66,0x92,0x00,0x10,0x00,0x00]
+          vpcmpgtd 4096(%rdx), %ymm22, %k2
+
+// CHECK: vpcmpgtd -4096(%rdx), %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf1,0x4d,0x20,0x66,0x52,0x80]
+          vpcmpgtd -4096(%rdx), %ymm22, %k2
+
+// CHECK: vpcmpgtd -4128(%rdx), %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf1,0x4d,0x20,0x66,0x92,0xe0,0xef,0xff,0xff]
+          vpcmpgtd -4128(%rdx), %ymm22, %k2
+
+// CHECK: vpcmpgtd 508(%rdx){1to8}, %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf1,0x4d,0x30,0x66,0x52,0x7f]
+          vpcmpgtd 508(%rdx){1to8}, %ymm22, %k2
+
+// CHECK: vpcmpgtd 512(%rdx){1to8}, %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf1,0x4d,0x30,0x66,0x92,0x00,0x02,0x00,0x00]
+          vpcmpgtd 512(%rdx){1to8}, %ymm22, %k2
+
+// CHECK: vpcmpgtd -512(%rdx){1to8}, %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf1,0x4d,0x30,0x66,0x52,0x80]
+          vpcmpgtd -512(%rdx){1to8}, %ymm22, %k2
+
+// CHECK: vpcmpgtd -516(%rdx){1to8}, %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf1,0x4d,0x30,0x66,0x92,0xfc,0xfd,0xff,0xff]
+          vpcmpgtd -516(%rdx){1to8}, %ymm22, %k2
+
+// CHECK: vpcmpgtq %xmm25, %xmm30, %k3
+// CHECK:  encoding: [0x62,0x92,0x8d,0x00,0x37,0xd9]
+          vpcmpgtq %xmm25, %xmm30, %k3
+
+// CHECK: vpcmpgtq %xmm25, %xmm30, %k3 {%k6}
+// CHECK:  encoding: [0x62,0x92,0x8d,0x06,0x37,0xd9]
+          vpcmpgtq %xmm25, %xmm30, %k3 {%k6}
+
+// CHECK: vpcmpgtq (%rcx), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x00,0x37,0x19]
+          vpcmpgtq (%rcx), %xmm30, %k3
+
+// CHECK: vpcmpgtq 291(%rax,%r14,8), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xb2,0x8d,0x00,0x37,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpgtq 291(%rax,%r14,8), %xmm30, %k3
+
+// CHECK: vpcmpgtq (%rcx){1to2}, %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x10,0x37,0x19]
+          vpcmpgtq (%rcx){1to2}, %xmm30, %k3
+
+// CHECK: vpcmpgtq 2032(%rdx), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x00,0x37,0x5a,0x7f]
+          vpcmpgtq 2032(%rdx), %xmm30, %k3
+
+// CHECK: vpcmpgtq 2048(%rdx), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x00,0x37,0x9a,0x00,0x08,0x00,0x00]
+          vpcmpgtq 2048(%rdx), %xmm30, %k3
+
+// CHECK: vpcmpgtq -2048(%rdx), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x00,0x37,0x5a,0x80]
+          vpcmpgtq -2048(%rdx), %xmm30, %k3
+
+// CHECK: vpcmpgtq -2064(%rdx), %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x00,0x37,0x9a,0xf0,0xf7,0xff,0xff]
+          vpcmpgtq -2064(%rdx), %xmm30, %k3
+
+// CHECK: vpcmpgtq 1016(%rdx){1to2}, %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x10,0x37,0x5a,0x7f]
+          vpcmpgtq 1016(%rdx){1to2}, %xmm30, %k3
+
+// CHECK: vpcmpgtq 1024(%rdx){1to2}, %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x10,0x37,0x9a,0x00,0x04,0x00,0x00]
+          vpcmpgtq 1024(%rdx){1to2}, %xmm30, %k3
+
+// CHECK: vpcmpgtq -1024(%rdx){1to2}, %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x10,0x37,0x5a,0x80]
+          vpcmpgtq -1024(%rdx){1to2}, %xmm30, %k3
+
+// CHECK: vpcmpgtq -1032(%rdx){1to2}, %xmm30, %k3
+// CHECK:  encoding: [0x62,0xf2,0x8d,0x10,0x37,0x9a,0xf8,0xfb,0xff,0xff]
+          vpcmpgtq -1032(%rdx){1to2}, %xmm30, %k3
+
+// CHECK: vpcmpgtq %ymm20, %ymm22, %k2
+// CHECK:  encoding: [0x62,0xb2,0xcd,0x20,0x37,0xd4]
+          vpcmpgtq %ymm20, %ymm22, %k2
+
+// CHECK: vpcmpgtq %ymm20, %ymm22, %k2 {%k1}
+// CHECK:  encoding: [0x62,0xb2,0xcd,0x21,0x37,0xd4]
+          vpcmpgtq %ymm20, %ymm22, %k2 {%k1}
+
+// CHECK: vpcmpgtq (%rcx), %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x20,0x37,0x11]
+          vpcmpgtq (%rcx), %ymm22, %k2
+
+// CHECK: vpcmpgtq 291(%rax,%r14,8), %ymm22, %k2
+// CHECK:  encoding: [0x62,0xb2,0xcd,0x20,0x37,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpcmpgtq 291(%rax,%r14,8), %ymm22, %k2
+
+// CHECK: vpcmpgtq (%rcx){1to4}, %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x30,0x37,0x11]
+          vpcmpgtq (%rcx){1to4}, %ymm22, %k2
+
+// CHECK: vpcmpgtq 4064(%rdx), %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x20,0x37,0x52,0x7f]
+          vpcmpgtq 4064(%rdx), %ymm22, %k2
+
+// CHECK: vpcmpgtq 4096(%rdx), %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x20,0x37,0x92,0x00,0x10,0x00,0x00]
+          vpcmpgtq 4096(%rdx), %ymm22, %k2
+
+// CHECK: vpcmpgtq -4096(%rdx), %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x20,0x37,0x52,0x80]
+          vpcmpgtq -4096(%rdx), %ymm22, %k2
+
+// CHECK: vpcmpgtq -4128(%rdx), %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x20,0x37,0x92,0xe0,0xef,0xff,0xff]
+          vpcmpgtq -4128(%rdx), %ymm22, %k2
+
+// CHECK: vpcmpgtq 1016(%rdx){1to4}, %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x30,0x37,0x52,0x7f]
+          vpcmpgtq 1016(%rdx){1to4}, %ymm22, %k2
+
+// CHECK: vpcmpgtq 1024(%rdx){1to4}, %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x30,0x37,0x92,0x00,0x04,0x00,0x00]
+          vpcmpgtq 1024(%rdx){1to4}, %ymm22, %k2
+
+// CHECK: vpcmpgtq -1024(%rdx){1to4}, %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x30,0x37,0x52,0x80]
+          vpcmpgtq -1024(%rdx){1to4}, %ymm22, %k2
+
+// CHECK: vpcmpgtq -1032(%rdx){1to4}, %ymm22, %k2
+// CHECK:  encoding: [0x62,0xf2,0xcd,0x30,0x37,0x92,0xf8,0xfb,0xff,0xff]
+          vpcmpgtq -1032(%rdx){1to4}, %ymm22, %k2
+
 // CHECK: vmovapd %xmm22, (%rcx)
 // CHECK:  encoding: [0x62,0xe1,0xfd,0x08,0x29,0x31]
           vmovapd %xmm22, (%rcx)
index 1d8561a..6ec147c 100644 (file)
@@ -217,6 +217,7 @@ static inline bool inheritsFrom(InstructionContext child,
   case IC_EVEX_KZ_B:
   case IC_EVEX_OPSIZE_K:
   case IC_EVEX_OPSIZE_B:
+  case IC_EVEX_OPSIZE_K_B:
   case IC_EVEX_OPSIZE_KZ:
     return false;
   case IC_EVEX_W_K:
@@ -224,11 +225,14 @@ static inline bool inheritsFrom(InstructionContext child,
   case IC_EVEX_W_XD_K:
   case IC_EVEX_W_OPSIZE_K:
   case IC_EVEX_W_OPSIZE_B:
+  case IC_EVEX_W_OPSIZE_K_B:
     return false;
   case IC_EVEX_L_K:
   case IC_EVEX_L_XS_K:
   case IC_EVEX_L_XD_K:
   case IC_EVEX_L_OPSIZE_K:
+  case IC_EVEX_L_OPSIZE_B:
+  case IC_EVEX_L_OPSIZE_K_B:
     return false;
   case IC_EVEX_W_KZ:
   case IC_EVEX_W_XS_KZ:
@@ -244,6 +248,8 @@ static inline bool inheritsFrom(InstructionContext child,
   case IC_EVEX_L_W_XS_K:
   case IC_EVEX_L_W_XD_K:
   case IC_EVEX_L_W_OPSIZE_K:
+  case IC_EVEX_L_W_OPSIZE_B:
+  case IC_EVEX_L_W_OPSIZE_K_B:
   case IC_EVEX_L_W_KZ:
   case IC_EVEX_L_W_XS_KZ:
   case IC_EVEX_L_W_XD_KZ:
index e4a40cb..a891d66 100644 (file)
@@ -1090,6 +1090,8 @@ RecognizableInstr::roRegisterEncodingFromString(const std::string &s,
   ENCODING("FR32X",           ENCODING_REG)
   ENCODING("VR512",           ENCODING_REG)
   ENCODING("VK1",             ENCODING_REG)
+  ENCODING("VK2",             ENCODING_REG)
+  ENCODING("VK4",             ENCODING_REG)
   ENCODING("VK8",             ENCODING_REG)
   ENCODING("VK16",            ENCODING_REG)
   ENCODING("VK32",            ENCODING_REG)